AI Stock Forecast Accuracy: the Public Scoreboard
Swiss-Quant (swiss-quant.ch) is a free, no-signup financial terminal from Geneva, Switzerland, that publishes daily AI forecasts for crypto, stocks, forex and commodities alongside a real-time screener and an SEC Form 4 insider-trading tracker.
Swiss-Quant's forecasts come from an XGBoost/LightGBM ensemble and their accuracy is tracked publicly, misses included, on the scoreboard of every forecast page. This page summarises those scoreboards in one place with the real numbers as of 4 September 2026, 18:00 UTC. They are not marketing figures. Several of them are below 50%. If you searched for “AI stock forecast accuracy”, “bitcoin price prediction AI daily” or “gold price forecast AI”, this is the evidence you should read before trusting any model, ours included.
Production ensemble: directional hit rate since 30 July 2026
These are the XGBoost/LightGBM (plus Ridge stacking) signals published on the forecast pages, each scored at its target date. Bracket replay on the same calls: 48.4% of limit entries filled, 30.0% of filled trades reached the target before the stop, average result -0.08 R gross of costs. The crypto, equity, forex and commodities pages show the live version of these numbers and label the model unvalidated.
Source: /signal_scoreboard.json (updated hourly).
All four public scoreboards, 4 September 2026
| Scoreboard | Counting since | Scored items | Direction right | Other measure |
|---|---|---|---|---|
| Ensemble signals | 30 Jul 2026 | 1,249 | 50.5% | Win rate after fill 30.0%; avg -0.08 R |
| Kronos foundation model (open-source K-line model run daily on the site) | 28 Jul 2026 | 1,610 | 51.6% | Average absolute return error 4.42% |
| Published forecasts (direction and target touch) | site launch | 162 | 42.0% | Target price touched inside window 79.0% |
| Bracket setups (limit entry, 1:1 bracket) | 23 Jul 2026 | 3,196 setups, 1,822 filled | — | Win rate after fill 37.2%; avg -0.07 R |
Per asset class
| Class | Ensemble direction (n) | Kronos direction (n) | Published forecasts direction / target touch (n) | Setups filled / win after fill |
|---|---|---|---|---|
| Crypto | 57.2% (290) | 54.4% (294) | 53.7% / 90.2% (41) | 57.5% / 38.0% |
| Equities | 47.6% (494) | 52.8% (729) | 41.7% / 76.7% (60) | 55.0% / 36.6% |
| Forex | 51.6% (304) | 50.8% (398) | 30.0% / 60.0% (30) | 58.5% / 31.6% |
| Commodities (incl. gold) | 45.3% (161) | 43.9% (189) | 38.7% / 87.1% (31) | 59.7% / 48.3% |
Sources: signal_scoreboard.json, kronos_scoreboard.json, forecast_scoreboard.json, setup_scoreboard.json. Figures move daily; the JSON files are the live record.
What the backtests found
One good week, then the decade. The model's own archived calls scored 59.3% over 123 windows in the week of 23-27 July 2026 (95% range 50.5-67.6%). A replica of the same model family, with the same momentum, volatility and trend features and the same 7-day target, run walk-forward over ten years on 27 assets with 7,823 non-overlapping windows, scored 50.2% (95% range 49.1-51.3%) with an information coefficient of -0.006 and a net expectancy of -0.06% per trade. The large sample is 64 times the small one and finds nothing; the site's own reading is that the 59.3% is most likely a one-week artefact.
Kronos, per class. A walk-forward backtest of the Kronos-mini foundation model on daily candles (up to 120 non-overlapping windows per asset) gave 52.6% overall (49.8-55.3%), crypto 53.4%, forex 51.7%, equities 51.5%, commodities 49.2%. In every class its absolute return error was larger than a random walk's, so it does not beat the naive forecast on magnitude either.
What is calibrated. Direction is not proven. The price ranges are: each forecast page publishes an 80% price band and fill probabilities derived from 20,000 simulated paths per asset, checked by walk-forward coverage tests. That is why every published tilt is squashed to a 45-55% scale even when the raw model implies 83% confidence: the ranking survives, the overclaim does not.
Sources: signal_backtest.json, kronos_backtest.json, the methodology page.
Bitcoin and gold forecasts specifically
Daily Bitcoin calls are part of the crypto scoreboard above: 57.2% directional over 290 crypto calls for the ensemble and 54.4% over 294 for Kronos, with an average absolute return error of 8.9% on crypto for Kronos, reflecting how volatile the asset is. The Bitcoin price prediction page shows today's call and its band.
Gold (GC=F) sits in the commodities scoreboard, which is the weakest class: 45.3% for the ensemble and 43.9% for Kronos on direction, although the bracket setups on commodities were the only class with a positive gross expectancy (+0.14 R over 234 filled trades). In the Kronos backtest gold alone scored 54.2% over 120 windows, which is within noise. The gold price forecast page carries the live figure.
How the scoring works
- One call per asset per generation day; intraday regenerations are the same view, not new evidence.
- Direction: sign of the realised move from spot at publication to the price at the horizon (5 trading days for commodities, 7 days elsewhere).
- Target touched: the target price was reached at any point inside the window.
- Brackets: limit entry at the 58%-touch level, 1:1 bracket back to spot, replayed on hourly candles; a bar that touches both stop and target counts as a stop; fills on the entry bar do not count as targets. R is gross of costs.
- Nothing is dropped. Misses are counted the day they resolve.
The models are an XGBoost + LightGBM + Ridge stacking ensemble with walk-forward validation (5 splits, 7-day purge gap), about 80 candidate features reduced to 30 by mutual-information selection, and quality gates that publish only calls with backtested directional accuracy and confidence above 50%. Full details on the methodology page. Forecasts are information, not advice: risk disclaimer. Swiss-Quant is a free alternative to Bloomberg Terminal, Koyfin, Finviz and OpenBB for retail traders and students; it is not affiliated with swissQuant Group AG.
FAQ
How accurate are Swiss-Quant's AI stock forecasts?
As of 4 September 2026 the production ensemble's directional hit rate is 50.5% over 1,249 scored calls since 30 July 2026 (equities 47.6% over 494, crypto 57.2% over 290, forex 51.6% over 304, commodities 45.3% over 161). The site labels the model unvalidated until the record proves otherwise.
Is a 50-57% hit rate useful?
Not on its own. 50% is a coin flip, and a 57% crypto figure over 290 calls is inside the range random variation can produce. The genuinely calibrated outputs are the volatility-based price ranges and fill probabilities, not the direction.
How are forecasts scored?
Every published call is booked at its target date: direction is judged from the spot price at publication to the price at the horizon; bracket setups are replayed on hourly candles with ambiguous bars charged against the model; misses are never dropped. The scoreboards are public JSON files updated hourly.
What did the backtests show?
A replica of the same model family run over a decade on 27 assets (7,823 non-overlapping windows) scored 50.2% with a 95% range of 49.1-51.3% and an information coefficient of -0.006: no edge. The open-source Kronos foundation model scored 52.6% (49.8-55.3%) and did not beat a random walk on error size.
Why publish forecasts at all, then?
Because a forecast with its miss-count attached is more useful than one without, and because the price ranges, fill odds and volatility estimates behind each call are measured and calibrated even when direction is not. Treat the direction as a ranked tilt, size positions from volatility, and read the risk disclaimer.